AI-102 vs DP-700: Which Certification Fits? : Cert-Pass Blog

Official source note

ai-102 vs dp-700 is the main focus of this page, and the safest way to study it is to keep both exam hubs open while you compare the goal of each certification and the service selection patterns that belong to each path. Microsoft describes Azure AI Engineer Associate and DP-700 Microsoft Fabric Data Engineer Associate as certifications that validate practical cloud literacy, service selection, and scenario thinking, but they reward different kinds of work. The main Cert Pass hubs remain /exams/azure-ai-102-azure-ai-engineer-associate and /exams/azure-dp-700-microsoft-fabric-data-engineer-associate.

Exam facts

Comparison topic: AI-102 vs DP-700
AI-102 hub: /exams/azure-ai-102-azure-ai-engineer-associate
DP-700 hub: /exams/azure-dp-700-microsoft-fabric-data-engineer-associate
AI-102 official page: Microsoft Azure AI Engineer Associate
DP-700 official page: Microsoft Fabric Data Engineer Associate

Why this comparison exists

The goal is not to decide which exam is globally better. The goal is to identify which exam fits the candidate's current role, target work, and preferred technical depth. If the goal is to build and evaluate AI solutions with Azure AI services, AI-102 is the better match. If the goal is to build Fabric data solutions, pipelines, and lakehouse style workflows, DP-700 is the better match.

Fast decision map

Use the two exam hubs side by side while you review: /exams/azure-ai-102-azure-ai-engineer-associate and /exams/azure-dp-700-microsoft-fabric-data-engineer-associate. Those internal links should act as the stable anchor for practice, revision, and final review.

Decision framework

If you want to focus on...	Choose...	Why
Azure AI services, LLM style solution design, prompt and model integration	AI-102	The exam centers on AI solution development in Azure
Fabric data engineering, lakehouse patterns, pipelines, and data platform work	DP-700	The exam centers on Microsoft Fabric data engineering
Application features that expose AI capabilities to users	AI-102	AI service integration is the main topic
Data ingestion, transformation, storage, and orchestration in Fabric	DP-700	Fabric data workflows are the main topic
Building AI powered apps and assistants	AI-102	The exam focuses on AI implementation patterns
End to end data engineering in Fabric	DP-700	The exam focuses on data platform implementation

AI-102 focused notes

The AI-102 path rewards candidates who understand how to select the right Azure AI service for a scenario, how to design an AI powered solution, and how to think about prompt, model, content, and integration choices in a practical way. A strong AI-102 candidate should be able to explain service selection, deployment boundaries, content safety, and the difference between using a ready service and designing a solution on top of it.

DP-700 focused notes

The DP-700 path rewards candidates who understand Fabric data engineering concepts, how to organize data for analytics, how to build and troubleshoot pipelines, and how to think about lakehouse and end to end platform workflows. A strong DP-700 candidate should be able to describe ingestion, transformation, monitoring, and the role of Fabric components in a data engineering solution.

Extended AI-102 revision notes

Microsoft AI-102 Exam Course: Designing and Implementing a Microsoft Azure AI Solution

Certification: Microsoft Certified: Azure AI Engineer Associate
Exam: AI-102: Designing and Implementing a Microsoft Azure AI Solution
Vendor: Microsoft

1. Exam Overview

What the exam is testing

AI-102 tests whether you can design, build, deploy, secure, monitor, and integrate Azure AI solutions. The exam is not only about knowing service names. Most questions ask you to choose the best service, architecture, API, deployment model, or security approach for a business scenario.

You should be ready to reason about:

Selecting the right Microsoft Foundry / Azure AI service for a requirement.
Building generative AI solutions with models, prompts, retrieval, evaluation, safety, and deployment.
Creating agentic solutions that use tools, grounding data, workflows, and guardrails.
Implementing computer vision, OCR, image analysis, video/image classification, and face-related scenarios.
Implementing language, speech, translation, summarization, sentiment, PII detection, and conversational workflows.
Building search, enrichment, document extraction, vector retrieval, and knowledge mining pipelines.
Applying Responsible AI, content safety, identity, network security, monitoring, cost control, and CI/CD.

How to think like the exam

The exam usually hides the answer in the requirement wording. Look for these clues:

Requirement clue	What it usually means
“Use one endpoint and key for multiple AI services”	Multi-service Azure AI services resource
“Least privilege, avoid keys”	Managed identity + RBAC
“Private network only”	Private endpoint, VNet integration, disable public network access where supported
“Extract fields from invoices/receipts/contracts/forms”	Azure AI Document Intelligence or Content Understanding, not generic OCR alone
“Search across documents with semantic/vector retrieval”	Azure AI Search with vector index, semantic ranking, skillsets where needed
“Ground model answers on enterprise data”	RAG with Azure AI Search or another retrieval layer
“Generate or evaluate prompts/workflows”	Microsoft Foundry tooling, prompt flow, evaluation, content safety
“Speech to text / text to speech / translation”	Azure AI Speech / Translator / Language depending on input and output
“Moderate harmful text/images”	Azure AI Content Safety
“Named entities, key phrases, sentiment, PII”	Azure AI Language

How to use this course

Read the domain overview first.
Study each domain by service-selection logic, not by memorizing isolated facts.
Use the tables to eliminate wrong answers quickly.
Review the traps section before practicing questions.
Use the final checklist as the last-day exam review.

2. Exam Domains

The official AI-102 domains are organized as follows. The CSV question bank was generated to match these priorities and then consolidated into this course.

Domain	Official priority	Rows in source CSV	Source share	What matters most
Plan and manage an Azure AI solution	20-25%	230	22.8%	Service selection, deployment, security, Responsible AI, monitoring, CI/CD
Implement generative AI solutions	15-20%	180	17.9%	Foundry, model deployment, prompts, RAG, grounding, evaluation, content safety
Implement an agentic solution	5-10%	75	7.4%	Agents, tools, grounding, orchestration, guardrails, evaluation
Implement computer vision solutions	10-15%	130	12.9%	Image analysis, OCR, custom vision concepts, face safety constraints
Implement natural language processing solutions	15-20%	190	18.8%	Language, Speech, Translator, PII, sentiment, summarization, conversational language
Implement knowledge mining and information extraction solutions	15-20%	203	20.1%	Azure AI Search, Document Intelligence, Content Understanding, indexing, vector search

Priority notes

High-yield areas from the source bank:

Security and deployment appear repeatedly across domains, especially managed identity, RBAC, Key Vault, private endpoints, and logging.
Azure AI Search is one of the most repeated services because it appears in knowledge mining, RAG, generative AI, and agentic grounding.
Document Intelligence is frequently contrasted with OCR, Azure AI Search, and generative AI.
Language, Speech, and Translator are commonly confused in scenario questions.
Foundry, generative AI, RAG, prompt evaluation, and content safety are emphasized in the latest blueprint.

3. Start-to-Finish Study Path

Foundation phase: know the service map

Study the purpose of each core service:

Microsoft Foundry / Azure AI Foundry: build, deploy, evaluate, and manage AI apps, models, prompt flows, and generative AI solutions.
Azure OpenAI / model deployment: chat, completions, embeddings, generative AI, summarization, classification, code/text generation.
Azure AI Search: lexical, semantic, hybrid, vector search, indexing, enrichment, retrieval for RAG.
Azure AI Document Intelligence: structured extraction from documents such as invoices, receipts, IDs, tax forms, and custom forms.
Azure AI Content Understanding: multimodal extraction and understanding across content types when the scenario goes beyond simple form extraction.
Azure AI Language: sentiment, key phrase extraction, named entity recognition, PII detection, summarization, classification, conversational language understanding.
Azure AI Speech: speech-to-text, text-to-speech, speech translation, pronunciation assessment, speaker-related capabilities where supported.
Azure AI Translator: text translation and document translation.
Azure AI Vision: OCR, image analysis, captions, object/tag detection, spatial/image insights where supported.
Azure AI Content Safety: detect harmful user or model-generated text/images.

Intermediate phase: learn decision rules

For each scenario, identify:

Input type: text, speech, image, video, document, mixed content.
Output type: extracted fields, search results, generated text, classification, translation, speech, summary.
Customization level: prebuilt model, custom model, prompt-based, fine-tuned/model deployment, custom extractor.
Security requirements: keyless auth, private network, data isolation, audit logging, compliance.
Integration style: SDK, REST API, container, managed endpoint, CI/CD pipeline, event-driven ingestion.
Operational needs: monitoring, evaluation, cost, latency, failover, versioning.

Advanced phase: master architecture tradeoffs

Practice these tradeoffs:

Prebuilt model vs custom model.
OCR vs Document Intelligence.
Azure AI Search indexing vs direct database query.
RAG vs fine-tuning vs prompt engineering.
Semantic search vs vector search vs hybrid search.
Single-service resource vs multi-service resource.
API key vs managed identity.
Public endpoint vs private endpoint.
Batch ingestion vs real-time inference.
Model evaluation vs application monitoring.

Final review phase

Before exam day, focus on:

Service selection tables.
Domain traps.
RAG architecture steps.
Azure AI Search indexing pipeline.
Document extraction pipeline.
Security and Responsible AI controls.
Prompt evaluation and content safety.
Difference between Language, Speech, Translator, Vision, Search, and Document Intelligence.

4. Core Concepts by Domain

Domain 1: Plan and manage an Azure AI solution

Concepts

This domain tests whether you can design and operate Azure AI solutions safely and correctly. It is not limited to provisioning resources. It includes choosing services, configuring deployment options, securing access, applying Responsible AI principles, monitoring, and integrating AI services into DevOps workflows.

Key concepts:

Selecting Microsoft Foundry Services based on task type.
Creating Azure AI resources and choosing single-service vs multi-service resources.
Choosing model deployment options and default endpoints.
Installing and using SDKs and REST APIs.
Securing secrets with Key Vault.
Using managed identities instead of hardcoded keys.
Applying RBAC and least privilege.
Configuring private endpoints and network restrictions.
Monitoring usage, latency, errors, quotas, and content safety events.
Building repeatable CI/CD deployment for AI apps.
Applying Responsible AI: fairness, reliability, safety, privacy, transparency, accountability.

Services

Service / capability	Use it for	Do not use it when
Multi-service Azure AI services resource	One endpoint/key for several AI services	You need per-service isolation, separate billing, or a service not supported by the multi-service resource
Single-service resource	Strong isolation, service-specific settings, separate quotas	The scenario requires one shared endpoint/key for many services
Managed identity	Keyless authentication to Azure resources	The target service does not support Entra ID/RBAC for that action
Key Vault	Store API keys, connection strings, secrets, certificates	Do not store secrets in app settings, source code, or notebooks
Private endpoint	Private network access to AI services	Public internet access is acceptable and simpler requirements are stated
Azure Monitor / Log Analytics	Metrics, logs, diagnostics, alerting	Do not use only application logs when platform metrics are required
Content Safety	Moderation and safety filters for user/model content	It does not replace identity, authorization, or network security

Patterns

Pattern: secure AI app integration

Recommended architecture:

Application uses managed identity.
Managed identity is granted minimum required RBAC role.
Secrets are stored in Key Vault only when keys are unavoidable.
AI service access is restricted with private endpoint if required.
Diagnostics are sent to Log Analytics.
Content Safety is applied where user-generated or model-generated content is involved.

Why wrong answers fail:

API keys in code are fast but insecure.
Storage account alone does not protect AI service calls.
Public endpoint with unrestricted keys fails private/compliance requirements.

Pattern: containerized AI service

Use containers when:

You need edge processing or low latency near data.
Connectivity is intermittent but billing/licensing requirements can still be met.
The selected service supports containers.

Trap: containers do not eliminate billing, licensing, connectivity, or service-specific limitations.

Traps

Choosing Azure Machine Learning when the scenario only needs a prebuilt AI API.
Choosing a multi-service resource when strict isolation or unsupported service features are required.
Choosing API keys when the requirement says “no secrets” or “least privilege.”
Assuming private endpoint automatically handles authorization; it only handles network path.
Confusing content moderation with security authorization.
Ignoring quotas, region availability, and model deployment constraints.

Domain 2: Implement generative AI solutions

Concepts

This domain tests your ability to build generative AI applications using Microsoft Foundry tooling, deployed models, prompt design, RAG, evaluation, and safety controls.

You must know:

Model selection and deployment.
Prompt engineering and system messages.
Grounding with enterprise data.
Retrieval-augmented generation.
Embeddings and vector search.
Prompt flow / orchestration patterns.
Evaluation of generated outputs.
Content safety and responsible AI guardrails.
Monitoring, cost, latency, and token usage.

Services

Capability	Best-fit service or pattern	Why
Chatbot over enterprise documents	Azure OpenAI/model deployment + Azure AI Search RAG	Keeps answers grounded in current internal content
Semantic/vector retrieval	Azure AI Search vector index	Enables similarity search over embeddings
Harmful content detection	Azure AI Content Safety	Detects unsafe user or model-generated content
Prompt workflow testing	Microsoft Foundry prompt flow/evaluation	Supports iterative prompt development and evaluation
Generate embeddings	Embedding model deployment	Converts text into vectors for retrieval
Need exact source citations	RAG with retrieved passages and metadata	Model alone cannot guarantee source grounding

Patterns

Pattern: RAG application

Use RAG when the question says:

“Answers must use company documents.”
“Information changes frequently.”
“Citations are required.”
“Do not retrain the model.”
“Use private knowledge base.”

Typical RAG flow:

Ingest documents into storage.
Chunk documents into meaningful sections.
Generate embeddings for chunks.
Store chunks, metadata, and vectors in Azure AI Search.
At query time, embed the user query.
Retrieve relevant chunks with vector, semantic, or hybrid search.
Pass retrieved context to the model.
Generate grounded answer with citations.
Apply content safety and logging.
Evaluate answer quality and retrieval quality.

Pattern: prompt safety and evaluation

Use:

System message to define role, boundaries, format, and refusal behavior.
Grounding data to reduce hallucination.
Content Safety to detect harmful inputs/outputs.
Evaluation datasets to compare prompt versions.
Logging and monitoring to detect failures in production.

Traps

Fine-tuning when RAG is better for changing knowledge.
Prompt engineering alone when source grounding is required.
RAG without chunking and metadata, causing poor retrieval.
Vector search alone when keyword matching is also important; hybrid search may be better.
Storing secrets in prompt flow or code instead of Key Vault.
Ignoring content safety for user-generated inputs.
Assuming the model can access private documents without retrieval integration.

Domain 3: Implement an agentic solution

Concepts

Agentic solutions use a model plus instructions, tools, functions, memory/context, retrieval, and evaluation to perform multi-step tasks. AI-102 focuses on safe, grounded, tool-using agents rather than uncontrolled autonomous systems.

You should understand:

Agent instructions and system prompts.
Tool/function calling.
Grounding with Azure AI Search or enterprise APIs.
Planning and step execution.
Human approval for sensitive actions.
Guardrails and tool permission boundaries.
Evaluation of agent task success.
Logging tool calls and outputs.

Services and capabilities

Requirement	Best pattern
Agent answers based on internal documents	Agent + RAG over Azure AI Search
Agent performs actions in business systems	Tool/function calling with scoped permissions
Agent must not execute destructive actions automatically	Human-in-the-loop approval
Agent must follow a specific workflow	Orchestrated prompt flow or explicit tool sequence
Agent outputs must be checked for harmful content	Content Safety and policy filters
Agent must be evaluated before production	Task-based evaluation with expected outputs and traces

Patterns

Pattern: safe tool-using agent

Define agent role and boundaries.
Register only required tools.
Give each tool minimal permissions.
Validate tool inputs and outputs.
Require approval for irreversible actions.
Use retrieval for factual grounding.
Log reasoning traces, tool calls, and final outputs where supported.
Evaluate with realistic task scenarios.

Traps

Giving an agent broad credentials when only read access is required.
Letting an agent write/delete/update records without approval.
Treating an agent like a simple chatbot when the scenario requires tool calls.
Using a generic model answer when the agent must call an API.
Ignoring retrieval quality; a smart agent with poor grounding still fails.
Failing to monitor tool-call errors and unsafe outputs.

Domain 4: Implement computer vision solutions

Concepts

This domain tests image and visual content processing. Focus on choosing between OCR, image analysis, custom vision scenarios, face capabilities, and document extraction services.

Key concepts:

Image analysis: tags, captions, object detection, visual features.
OCR: extract printed/handwritten text from images.
Document extraction: use Document Intelligence for structured forms/documents.
Custom image classification/object detection: train custom models where supported.
Face capabilities: identity/sensitive use cases require careful Responsible AI and service restrictions.
Batch vs real-time image processing.
Containers for edge vision where supported.

Services

Scenario	Use	Avoid
Read text from a simple image	Azure AI Vision OCR	Full Document Intelligence unless structured fields/forms are required
Extract invoice fields like vendor, total, date	Azure AI Document Intelligence prebuilt invoice model	Generic OCR only
Generate image tags/captions	Azure AI Vision image analysis	Azure AI Language
Detect harmful images	Azure AI Content Safety	Azure AI Vision alone
Custom image classifier/object detector	Custom vision/model option where current service supports it	Generic image analysis if custom labels are needed
Analyze ID/document with structured fields	Document Intelligence	Vision OCR only

Patterns

Pattern: simple OCR vs structured document extraction

If the scenario only asks for text extraction from an image, OCR is enough.
If the scenario asks for fields, tables, key-value pairs, layout, invoices, receipts, contracts, or forms, use Document Intelligence.

Pattern: image moderation

Image analysis identifies objects/tags/captions.
Content Safety assesses harmful or unsafe content.
For user uploads, moderation often happens before storage, indexing, or model processing.

Traps

Choosing OCR for invoice field extraction.
Choosing Document Intelligence when the scenario only needs image tags.
Choosing Language service for image analysis.
Ignoring Responsible AI restrictions in face-related scenarios.
Assuming every Vision capability is available in every region or deployment mode.

Domain 5: Implement natural language processing solutions

Concepts

This domain covers text and speech understanding. The biggest exam skill is selecting the correct service for the input/output type.

You need to know:

Sentiment analysis and opinion mining.
Key phrase extraction.
Named entity recognition.
PII detection and redaction.
Text summarization.
Language detection.
Custom text classification.
Conversational language understanding.
Question answering where applicable.
Speech-to-text and text-to-speech.
Translation and document translation.

Services

Scenario	Best service
Detect sentiment in customer reviews	Azure AI Language sentiment analysis
Extract organizations, people, locations	Azure AI Language named entity recognition
Detect and redact personal data	Azure AI Language PII detection
Extract key phrases from text	Azure AI Language key phrase extraction
Classify support tickets by custom labels	Azure AI Language custom text classification
Understand user intent in a bot	Conversational language understanding
Convert audio to text	Azure AI Speech speech-to-text
Convert text to natural voice	Azure AI Speech text-to-speech
Translate text between languages	Azure AI Translator
Translate audio conversations	Speech translation, not text-only Translator

Patterns

Pattern: support-ticket routing

Use Azure AI Language custom text classification when:

The input is text.
You need labels such as Billing, Technical Support, Cancellation, Complaint.
You have training examples for custom classes.

Use conversational language understanding when:

You need to detect intents and entities from interactive user utterances.
The input is part of a bot or conversation.

Pattern: compliance redaction

Use Language PII detection when:

You need to find names, emails, phone numbers, addresses, government IDs, or other personal data.
You need redaction before storage, logging, or downstream model calls.

Traps

Choosing Translator when the requirement is sentiment or PII detection.
Choosing Speech for text-only translation.
Choosing Language when the input is audio and no transcription has occurred.
Choosing generic generative AI for deterministic PII detection when a dedicated Language capability exists.
Confusing entity extraction with key phrase extraction.
Confusing text classification with conversational intent recognition.

Domain 6: Implement knowledge mining and information extraction solutions

Concepts

This domain is heavily tested because it combines search, enrichment, document extraction, indexing, and retrieval for generative AI. You must understand how content becomes searchable and how structured data is extracted from unstructured sources.

Key concepts:

Azure AI Search indexes.
Data sources, indexers, and skillsets.
Built-in skills: OCR, entity recognition, key phrase extraction, translation, language detection, image analysis.
Custom skills for external enrichment.
Knowledge stores/projections where applicable.
Semantic search, vector search, hybrid search.
Scoring profiles, filters, facets, analyzers, synonym maps.
Document Intelligence models and extraction results.
Content Understanding for multimodal extraction workflows.
RAG retrieval patterns.

Services

Requirement	Best choice
Make many documents searchable	Azure AI Search index
Enrich documents during indexing	Azure AI Search skillset
Extract structured fields from invoices/forms	Azure AI Document Intelligence
Extract searchable text from images in documents	OCR skill or Document Intelligence depending on structure
Add custom enrichment logic	Custom skill in Azure AI Search pipeline
Search by meaning rather than exact words	Vector search or semantic search
Combine keyword and vector relevance	Hybrid search
Ground generative AI answers	Azure AI Search retrieval + model context

Patterns

Pattern: indexing pipeline

Store documents in a supported source such as Azure Blob Storage.
Create a data source connection.
Define an index schema with searchable, filterable, facetable, sortable fields.
Define an indexer.
Add a skillset if enrichment is required.
Add vector fields if vector retrieval is required.
Run the indexer and monitor errors/warnings.
Query with filters, semantic ranking, vector search, or hybrid search.

Pattern: document extraction pipeline

Receive document.
Choose prebuilt or custom Document Intelligence model.
Extract fields/tables/layout/key-value pairs.
Validate confidence scores.
Store structured results.
Optionally index extracted text/metadata in Azure AI Search.
Use results in downstream workflows or RAG.

Traps

Choosing Azure AI Search to extract invoice fields directly; Search indexes content, while Document Intelligence extracts structured document fields.
Choosing Document Intelligence to rank search results; that is Azure AI Search.
Using only vector search when exact filters/facets are required.
Forgetting to mark fields as filterable/facetable/sortable at index design time.
Forgetting skillset output mappings.
Expecting an indexer to understand unsupported file formats without enrichment or custom logic.
Ignoring indexer error monitoring.

5. Service Selection Guide

Core AI service decision table

If the scenario says...	Choose...	Because...
Generate text, summarize, reason, chat	Azure OpenAI/model deployment in Foundry	Generative model capability
Answer from private documents	RAG with Azure AI Search	Grounds answers in enterprise content
Search documents by keyword, semantic meaning, or vector similarity	Azure AI Search	Search/retrieval platform
Extract fields from documents	Document Intelligence	Purpose-built document extraction
Understand multimodal content	Content Understanding	Works across content types and extraction scenarios
Analyze image content	Azure AI Vision	Tags, captions, OCR, image features
Moderate harmful content	Content Safety	Safety classification for text/images
Analyze text sentiment, entities, PII, key phrases	Azure AI Language	NLP extraction/classification
Translate text	Translator	Text translation
Convert speech/audio	Speech	Speech-to-text, text-to-speech, speech translation
Build tool-using assistant	Agentic solution with tools, RAG, guardrails	Multi-step task execution

Confusing service comparisons

Azure AI Search vs Document Intelligence

Feature	Azure AI Search	Document Intelligence
Main purpose	Index and search content	Extract structured information from documents
Output	Search results, rankings, facets, snippets	Fields, tables, key-value pairs, layout, confidence scores
Typical input	Documents, metadata, extracted text	PDFs, images, forms, invoices, receipts, contracts
Used in RAG	Yes, as retriever	Sometimes, to extract content before indexing
Common trap	Treating it as an extraction engine	Treating it as a search engine

Azure AI Language vs Azure OpenAI

Requirement	Azure AI Language	Azure OpenAI / generative model
Deterministic PII detection	Strong fit	Not preferred as primary control
Sentiment analysis	Strong fit	Possible but less specialized
Custom text classification	Strong fit	Possible with prompt/fine-tune, but not first choice for classic classification
Open-ended generation	Not the purpose	Strong fit
Summarization	Supported in Language and generative models depending on scenario	Strong fit when flexible generation is needed
Need strict safety/moderation	Pair with Content Safety	Pair with Content Safety

Translator vs Speech

Requirement	Choose
Translate text to another language	Translator
Translate documents	Translator document translation
Convert audio to text	Speech-to-text
Convert text to voice	Text-to-speech
Translate spoken audio	Speech translation

RAG vs fine-tuning vs prompt engineering

Pattern	Use when	Avoid when
Prompt engineering	You need better instructions, format, tone, or constraints	You need private/current facts not in the model
RAG	You need grounded answers from changing enterprise data	You need to teach a model a stable task style or domain behavior only
Fine-tuning	You need consistent style, task behavior, or domain-specific patterns	You need frequently changing facts or citations

6. Architecture Patterns

Scenario 1: Enterprise document chatbot

Recommended solution: Azure OpenAI/model deployment + Azure AI Search + embeddings + Content Safety + managed identity.

Why:

Azure AI Search retrieves trusted document chunks.
Embeddings enable semantic similarity.
The model generates grounded answers.
Content Safety checks harmful inputs/outputs.
Managed identity avoids hardcoded keys.

Wrong alternatives:

Fine-tuning alone: does not keep current documents fresh.
Prompt only: cannot access private data.
Document Intelligence only: extracts content but does not provide full chatbot retrieval and generation.

Scenario 2: Invoice processing workflow

Recommended solution: Document Intelligence prebuilt invoice model or custom model, then store extracted fields and optionally index results.

Why:

Invoices require structured extraction: vendor, date, total, line items, tax.
OCR alone gives text, not reliable field mapping.
Azure AI Search can index extracted results but is not the extraction model.

Scenario 3: Secure banking chatbot

Recommended solution: private endpoint, managed identity, Key Vault only for unavoidable secrets, Content Safety, audit logs, PII detection/redaction where needed, and RAG over approved sources.

Wrong alternatives:

Public endpoint with unrestricted API key.
Logs containing raw PII.
Sending unredacted sensitive data to downstream services without controls.

Scenario 4: Customer support ticket automation

Recommended solution: Azure AI Language for classification, sentiment, PII detection; optionally generative AI for response drafting; human approval before sending sensitive responses.

Wrong alternatives:

Translator if no translation is needed.
Speech if the input is text.
Generative AI alone for compliance-grade PII detection.

Scenario 5: Knowledge mining from PDFs and scanned images

Recommended solution: Azure Blob Storage + Azure AI Search indexer + skillset with OCR/entity/key phrase skills + optional Document Intelligence + index fields designed for filters/facets + monitoring.

Wrong alternatives:

Search index without enrichment for scanned images.
Missing field attributes such as filterable/facetable.
Ignoring skillset output mappings.

Scenario 6: Agent that books appointments

Recommended solution: agent with clear instructions, calendar/tool APIs, least-privilege credentials, validation, human confirmation for booking/cancellation, and logs.

Wrong alternatives:

Chat model with no tools cannot actually book.
Agent with full admin rights violates least privilege.
No approval for destructive or external actions is risky.

7. Exam Traps

Misleading wording patterns

Wording	Trap	Correct thinking
“AI solution”	May tempt you to choose Azure ML	Use Azure AI services when prebuilt APIs satisfy the requirement
“Search documents”	May tempt Document Intelligence	Search/ranking is Azure AI Search
“Extract invoice total”	May tempt OCR	Structured document fields need Document Intelligence
“Translate spoken audio”	May tempt Translator	Use Speech translation
“Analyze image safety”	May tempt Vision	Use Content Safety for harmful content
“No secrets in code”	May still show API key options	Use managed identity/RBAC where supported
“Current company policies”	May tempt fine-tuning	Use RAG because facts change
“Private access”	May tempt Key Vault only	Use private endpoint/network controls plus identity

Wrong-but-plausible answers

Azure Machine Learning appears as a distractor when a prebuilt Azure AI service is enough.
Azure Storage appears as a distractor for security or AI processing but is only storage.
API keys appear as easy integration choices but fail least-privilege/no-secret requirements.
OCR appears as a distractor for document understanding; use it only when raw text is enough.
Fine-tuning appears as a distractor for private knowledge; RAG is usually better.
Translator appears in speech scenarios; Speech is needed when audio is involved.
Azure AI Search appears for field extraction; it indexes and retrieves, it does not replace extraction models.

Elimination strategy

When stuck, ask:

What is the input format?
What exact output is required?
Does the scenario require extraction, search, generation, classification, or translation?
Does the answer satisfy security wording such as private, least privilege, or no secrets?
Does the solution use the simplest service that directly satisfies the requirement?
Does the data change frequently? If yes, prefer retrieval over fine-tuning.
Is the action risky? If yes, require guardrails and human approval.

8. Quick Memory Rules

Rules of thumb

Forms/invoices/receipts/contracts → Document Intelligence.
Search/retrieval/RAG → Azure AI Search.
Generated answers over private data → RAG.
Changing facts → RAG, not fine-tuning.
Style/behavior consistency → prompt engineering or fine-tuning.
PII/sentiment/entities/key phrases → Azure AI Language.
Audio in or audio out → Speech.
Text translation → Translator.
Image tags/captions/OCR → Vision.
Unsafe text/images → Content Safety.
No secrets → managed identity + RBAC.
Private traffic → private endpoint.
Secrets unavoidable → Key Vault.
Multiple AI services, one endpoint → multi-service Azure AI services resource.

“If you see X, think Y” patterns

If you see...	Think...
“Grounded on internal documents”	RAG + Azure AI Search
“Citations required”	Retrieve passages with metadata
“Hallucination reduction”	Grounding, prompt constraints, evaluation, content safety
“Evaluate prompt quality”	Foundry evaluation / prompt flow evaluation
“Extract tables from PDFs”	Document Intelligence
“Scanned PDFs searchable”	OCR/enrichment + Azure AI Search
“Filter by department/date/category”	Search index fields must be filterable
“Facet by product/category”	Search index fields must be facetable
“Vector similarity”	Embeddings + vector index
“Hybrid relevance”	Keyword + vector + optional semantic ranking
“Detect customer intent”	Conversational language understanding
“Route support tickets”	Custom text classification
“Redact emails/phone numbers”	PII detection
“Speech pronunciation scoring”	Speech pronunciation assessment
“Moderate prompts and completions”	Content Safety
“Edge/intermittent connectivity”	Containers, but check service support and billing requirements

9. Final Revision Notes

Highest-yield review points

Service selection is the core exam skill. Always match input and output.
Security wording is decisive. No secrets means managed identity; private means private endpoint; secrets mean Key Vault.
RAG is the default for private/changing knowledge. Fine-tuning is not a knowledge database.
Azure AI Search powers retrieval. It is central to knowledge mining and generative AI grounding.
Document Intelligence extracts structured document data. Do not replace it with OCR when fields/tables are required.
Language, Speech, Translator, and Vision are separated by input type. Text, audio, translation, and image requirements point to different services.
Agents need tools and guardrails. A model without tools cannot perform external actions.
Responsible AI is practical. Use safety filters, monitoring, human review, and evaluation.
Index design matters. Search fields must be configured correctly before they can be filtered, sorted, faceted, or vectorized.
Exam answers prefer managed Azure services over custom code unless the scenario explicitly requires custom logic.

Last-day revision list

Review all six official domains and weightings.
Memorize the service selection table.
Practice distinguishing OCR, Document Intelligence, Content Understanding, and Azure AI Search.
Practice RAG architecture from ingestion to answer generation.
Review identity and network security patterns.
Review Language vs Translator vs Speech.
Review content safety and Responsible AI controls.
Review agentic solution patterns and human approval.
Review Azure AI Search field attributes, skillsets, vector fields, semantic ranking, and indexers.
Review prompt evaluation and monitoring.

10. Exam-Day Checklist

Must-know topics

Official AI-102 domains and priorities.
Microsoft Foundry service-selection logic.
Multi-service vs single-service Azure AI resources.
Managed identity, RBAC, Key Vault, private endpoint.
Responsible AI principles and practical controls.
Content Safety for text and image moderation.
Generative AI model deployment and prompt design.
RAG with embeddings and Azure AI Search.
Prompt flow/evaluation concepts.
Agentic solutions with tools, grounding, logging, and approval.
Vision OCR vs image analysis vs document extraction.
Document Intelligence prebuilt/custom models.
Azure AI Language: sentiment, NER, PII, key phrases, classification, summarization.
Speech: speech-to-text, text-to-speech, speech translation.
Translator: text and document translation.
Azure AI Search: indexes, indexers, skillsets, vector search, semantic search, filters/facets.
Knowledge mining enrichment pipeline.
Monitoring, diagnostics, quotas, latency, cost, and deployment automation.

Final confidence checklist

Before submitting an answer, confirm:

The selected service matches the required input and output.
The solution satisfies all security requirements.
The architecture avoids unnecessary custom development.
The answer does not use a service for the wrong purpose.
If facts change often, the answer uses retrieval instead of training.
If structured document fields are needed, the answer uses Document Intelligence.
If search/ranking/retrieval is needed, the answer uses Azure AI Search.
If harmful content is involved, the answer includes Content Safety.
If an agent takes actions, the answer includes scoped tools and guardrails.
If the scenario says “least privilege,” the answer avoids broad keys and permissions.

Compact Final Map

Need	Best answer
One endpoint for multiple AI services	Multi-service Azure AI services resource
Private document chatbot	RAG + Azure AI Search + model deployment
Extract invoice fields	Document Intelligence
Make scanned PDFs searchable	OCR/enrichment + Azure AI Search
Detect PII	Azure AI Language
Translate text	Translator
Process audio	Speech
Analyze images	Vision
Moderate harmful content	Content Safety
Avoid secrets	Managed identity + RBAC
Store unavoidable secrets	Key Vault
Private connectivity	Private endpoint
Execute external actions with AI	Agent + tools + guardrails
Evaluate prompt/model output	Foundry evaluation/prompt flow

End of course.

Extended DP-700 revision notes

DP-700 Microsoft Fabric Data Engineer Associate - Compressed Exam Course

Built from the provided practice CSV/question bank (1050 questions) and consolidated into original revision notes. The source bank is evenly distributed across the three DP-700 domains: Implement and manage an analytics solution: 350 questions, Ingest and transform data: 350 questions, Monitor and optimize an analytics solution: 350 questions. Use this file as a fast, scenario-focused study guide, not as a question-by-question summary.

1. Exam Overview

What the exam is testing

DP-700 validates whether you can implement data engineering solutions in Microsoft Fabric. The exam is not just about knowing product names. It tests whether you can choose the right Fabric item, loading pattern, transformation engine, security model, monitoring approach, and optimization technique for a realistic enterprise analytics scenario.

You are expected to reason across:

Workspaces and lifecycle: Git integration, deployment pipelines, environments, item promotion, workspace settings, domains, capacity, and governance.
Data engineering implementation: lakehouses, warehouses, Eventhouses, Eventstreams, Dataflows Gen2, notebooks, pipelines, KQL, T-SQL, PySpark, shortcuts, mirroring, batch and streaming ingestion.
Operations and performance: troubleshooting pipelines, notebooks, Dataflows Gen2, Eventstreams, Eventhouses, OneLake shortcuts, semantic model refresh, Spark jobs, warehouse queries, and capacity issues.

How to think like the exam

The exam usually gives you a business or technical constraint and asks for the best Fabric-native choice. Do not choose the tool you personally prefer. Choose the tool that best matches the scenario constraints.

Typical exam logic:

Identify the data shape: batch, streaming, relational, files, telemetry, dimensional model, or operational replication.
Identify the user persona: data engineer, low-code analyst, SQL developer, real-time analyst, BI consumer, administrator.
Identify operational constraints: CI/CD, governance, security, monitoring, cost, performance, incremental load, late-arriving data, or schema evolution.
Eliminate attractive but wrong options: wrong engine, wrong security layer, wrong optimization level, or manual approach when Fabric has a managed feature.
Prefer the simplest Fabric-native solution that satisfies all requirements.

How to use this course

Read sections 1-3 first, then study sections 4-8 by scenario. For final review, use sections 9-10. When practicing questions, map every question to one of these decisions:

Which Fabric item should be used?
Which transformation engine is best?
Which security boundary applies?
Which monitoring signal identifies the problem?
Which optimization action fixes the bottleneck?

2. Exam Domains

Official domain	Weight	What matters most	Source-bank emphasis
Implement and manage an analytics solution	30-35%	Workspace settings, lifecycle management, security, governance, orchestration	350 questions
Ingest and transform data	30-35%	Batch and streaming ingestion, transformation engines, loading patterns, OneLake, shortcuts, mirroring	350 questions
Monitor and optimize an analytics solution	30-35%	Monitoring, troubleshooting, semantic refresh, pipeline/notebook/Eventhouse errors, performance tuning	350 questions

Priority notes

All three DP-700 domains have similar weights. The practical priority is:

Ingest and transform data - this is where many scenario questions hide the service-selection decision.
Implement and manage analytics solutions - governance, CI/CD, access control, and orchestration are frequent traps.
Monitor and optimize analytics solutions - questions often test the exact diagnostic surface or optimization action.

What matters most

Know how to distinguish these pairs quickly:

Dataflow Gen2 vs notebook vs pipeline vs T-SQL vs KQL.
Lakehouse vs warehouse vs Eventhouse.
Shortcut vs copy vs mirroring.
Full load vs incremental load vs streaming load.
Workspace role vs item permission vs OneLake security vs SQL security.
Deployment pipeline vs Git integration.
Pipeline failure vs notebook failure vs Dataflow Gen2 refresh failure vs semantic model refresh failure.
Spark optimization vs warehouse query optimization vs Eventhouse/KQL optimization.

3. Start-to-Finish Study Path

Foundation: understand the Fabric data platform

Start with the Fabric object model:

Workspace: collaboration and security boundary for Fabric items.
OneLake: tenant-wide data lake foundation.
Lakehouse: file/table-oriented engineering store backed by Delta tables and Spark.
Warehouse: relational SQL analytics store for T-SQL developers and dimensional workloads.
Eventhouse: real-time analytics store optimized for event/telemetry data and KQL.
Data pipeline: orchestration, movement, scheduling, dependencies, parameters.
Dataflow Gen2: low-code/no-code Power Query-based ingestion and transformation.
Notebook: PySpark/SQL code-first transformation and engineering.
Eventstream: real-time event ingestion and routing.

Foundation goal: when you see a requirement, you should immediately know the most likely Fabric item.

Intermediate: master ingestion and transformation decisions

Study these loading patterns:

Full load for small or replaceable data.
Incremental load with watermark for large changing data.
Change data capture or mirroring when operational replication is required.
Streaming ingestion for continuous events.
Bronze/Silver/Gold pattern for lakehouse engineering.
Dimensional modeling preparation for warehouse or BI consumption.

Intermediate goal: explain why one engine is better than another for a given scenario.

Advanced: governance, CI/CD, orchestration, and reliability

Focus on:

Git integration for version control and pull-request workflows.
Deployment pipelines for controlled promotion across dev/test/prod.
Workspace roles and item permissions.
Row-level, column-level, object-level, folder/file-level, and OneLake security.
Sensitivity labels and endorsement.
Fabric audit logs.
Pipelines with parameters, dynamic expressions, retries, schedules, and event triggers.

Advanced goal: design a production-ready solution, not just a working data load.

Final review: monitoring and optimization

Practice recognizing symptoms:

Slow Spark notebook: partitioning, shuffle, skew, file size, caching, job metrics.
Slow warehouse query: statistics, distribution of joins, indexing/physical design where applicable, query plan, materialization strategy.
Lakehouse table issue: Delta maintenance, compaction, vacuum retention, file layout.
Pipeline failure: activity output, dependency, parameter, linked connection, schema drift, permission.
Eventstream/Eventhouse issue: ingestion errors, schema mapping, retention, KQL function/windowing, throughput.

Final goal: when a question describes a failure, know where to look first and which fix is targeted.

4. Core Concepts by Domain

Domain 1: Implement and manage an analytics solution

Concepts

This domain tests whether you can configure and manage Fabric solutions as enterprise assets. It is not only about creating lakehouses or notebooks; it is about controlling how they are secured, promoted, governed, and orchestrated.

Key concepts:

Workspace configuration for Spark, domains, OneLake, and Dataflows Gen2.
Version control and collaboration with Git integration.
Controlled deployment with deployment pipelines.
Database projects for warehouse development lifecycle.
Workspace-level and item-level access control.
SQL security and OneLake security.
Sensitivity labels, endorsement, and audit logs.
Orchestration with pipelines, notebooks, parameters, dynamic expressions, schedules, and event triggers.

Services

Need	Best Fabric choice	Why
Branching, pull requests, rollback	Git integration	Source-control workflow for collaboration and change history
Promote items from dev to test to prod	Deployment pipeline	Environment promotion, comparison, deployment rules
Schedule multi-step workloads	Data pipeline	Orchestration, dependencies, parameters, retry logic
Run complex code transformations	Notebook	PySpark/SQL code, reusable logic, engineering flexibility
Low-code transformation	Dataflow Gen2	Power Query experience and managed refresh
Govern data classification	Sensitivity labels	Applies classification and protection metadata
Certify trusted assets	Endorsement	Helps users identify promoted/certified content
Investigate user/admin activity	Audit logs	Trace actions and governance events

Patterns

Use Git integration for developer collaboration; use deployment pipelines for release promotion.
Use workspace roles for broad collaboration access; use item permissions for specific artifacts.
Use SQL row/column/object-level security for SQL access patterns; use OneLake security for file/folder/table access patterns in OneLake.
Use pipelines as the orchestrator and call notebooks, Dataflows Gen2, copy activities, or stored procedures as steps.
Use parameters and dynamic expressions to avoid hardcoding paths, dates, workspace names, and environment values.

Traps

Choosing Git integration when the requirement is environment promotion and approvals. Correct answer is usually deployment pipeline.
Choosing deployment pipeline when the requirement is pull requests and branch history. Correct answer is usually Git integration.
Choosing workspace Admin when the user only needs to read one item. Prefer least privilege.
Applying sensitivity labels when the requirement is to restrict rows. Sensitivity labels classify; they do not replace row-level security.
Using a notebook as the orchestrator when the requirement is scheduling, dependency management, retries, and monitoring. Pipelines are usually the orchestrator.

Domain 2: Ingest and transform data

Concepts

This is the largest practical part of the exam because it tests service selection. The same data can often be transformed by Dataflows Gen2, notebooks, T-SQL, KQL, or pipelines. The exam wants the best fit.

Key concepts:

Full, incremental, and streaming loading patterns.
Watermark-based incremental ingestion.
Dimensional model preparation.
Lakehouse, warehouse, and Eventhouse selection.
OneLake shortcuts versus physical copy.
Mirroring for operational data replication.
Batch ingestion with pipelines.
Transformations using PySpark, SQL, and KQL.
Handling duplicates, missing values, and late-arriving data.
Eventstreams, Spark structured streaming, KQL processing, and windowing functions.

Services

Need	Best choice	Why
Large-scale file/table transformation	Notebook with Spark	Scalable, code-first, complex transformations
Low-code ingestion/transformation	Dataflow Gen2	Power Query, accessible for analysts, managed refresh
SQL transformation in warehouse	T-SQL	Relational logic, dimensional models, SQL developer workflow
Real-time telemetry analysis	Eventhouse + KQL	Optimized for event/time-series analytics
Real-time ingestion/routing	Eventstream	Event capture, routing, filtering, stream processing entry point
Orchestrate copy and transformations	Pipeline	Scheduling and dependencies across steps
Access data without copying	OneLake shortcut	Virtual access to data in another location
Replicate operational data	Mirroring	Near real-time replication with less custom ETL
Handle continuously arriving data in Spark	Spark structured streaming	Code-based stream processing

Patterns

Use watermarks for incremental batch loads. Store the last successful load timestamp or key.
Use deduplication keys and event time when duplicate or late-arriving records are possible.
Use Eventstream to ingest and route events; use Eventhouse/KQL to query and analyze event data.
Use shortcuts when data should remain in place and be accessed through OneLake.
Use copy/movement when you need physical control, transformation during landing, or isolation from source changes.
Use mirroring when the requirement is operational database replication into Fabric with minimal ETL.
Use lakehouse for engineering and open data layout; use warehouse for SQL-first curated analytics and dimensional modeling.

Traps

Choosing a warehouse for raw semi-structured file engineering when a lakehouse/notebook pattern fits better.
Choosing a notebook for simple low-code transformation when Dataflow Gen2 is enough and maintainable by analysts.
Choosing Dataflow Gen2 for very complex PySpark logic, custom libraries, or distributed code workflows. Use notebooks.
Choosing a shortcut when the requirement says transform and store a curated copy. Shortcut is access, not transformation.
Choosing full load for large frequently changing data. Incremental with watermark is preferred.
Ignoring late-arriving data in streaming questions. Use event-time windowing and proper watermarking logic.

Domain 3: Monitor and optimize an analytics solution

Concepts

This domain tests operational judgment. The exam often describes symptoms and asks what you should inspect or optimize.

Key concepts:

Monitoring ingestion, transformation, and semantic model refresh.
Pipeline run history, activity output, retries, and dependency diagnostics.
Dataflow Gen2 refresh errors and transformation-step issues.
Notebook execution errors, Spark job metrics, logs, and resource bottlenecks.
Eventstream and Eventhouse ingestion/query errors.
T-SQL error diagnosis and warehouse query tuning.
OneLake shortcut errors caused by path, permission, source availability, or schema issues.
Lakehouse table optimization, compaction, vacuuming, and query layout.
Spark performance tuning: partitions, skew, shuffle, caching, file sizes.
Warehouse and KQL query optimization.

Services and diagnostics

Symptom	First place to inspect	Likely fix
Pipeline activity failed	Pipeline run details and activity output	Correct parameter, connection, dependency, schema, or permission
Notebook runs slowly	Spark UI/job metrics/logs	Reduce shuffle, repartition, handle skew, cache selectively
Lakehouse table has many small files	Lakehouse/Delta optimization tools	Compact/optimize table and manage retention carefully
Dataflow Gen2 refresh fails	Dataflow refresh history and step errors	Fix transformation step, schema mismatch, credentials, or destination mapping
Semantic model refresh fails	Refresh history and data source credentials	Fix credentials, gateway/connection, capacity, or upstream data availability
Eventhouse ingestion fails	Ingestion diagnostics and mappings	Fix schema mapping, format, batching, retention, or permission
KQL query slow	Query diagnostics and KQL design	Filter early, reduce scanned data, use time filters, summarize efficiently
Warehouse query slow	Query plan/performance view	Reduce scans, improve joins, update statistics/materialize where appropriate
Shortcut broken	Shortcut target and permissions	Fix source path, credentials, permissions, or source availability

Patterns

Diagnose before optimizing. The exam often rewards the answer that checks the specific run details or metrics first.
For Spark, think: shuffle, partitions, skew, cache, file size.
For lakehouse Delta tables, think: optimize/compact, vacuum carefully, partition wisely.
For streaming, think: throughput, schema mapping, event-time windows, late data, retention.
For pipelines, think: activity output, dependencies, retry policy, parameters, connections.
For semantic model refresh, think: upstream availability, credentials, capacity, refresh history.

Traps

Restarting capacity before checking run-level diagnostics. Capacity can be relevant, but exam questions often expect targeted troubleshooting first.
Vacuuming as a universal fix. Vacuum removes old files; it can break time travel if retention is too aggressive.
Partitioning by high-cardinality columns. It can create too many small files.
Caching everything in Spark. Cache only reused intermediate data; otherwise it wastes memory.
Optimizing the wrong layer: Spark tuning will not fix a SQL warehouse query plan problem, and warehouse tuning will not fix Eventhouse ingestion mapping.

5. Service Selection Guide

Lakehouse vs Warehouse vs Eventhouse

Requirement	Lakehouse	Warehouse	Eventhouse
Primary persona	Data engineers, Spark users	SQL developers, BI/analytics engineers	Real-time analytics engineers
Best for	Files, Delta tables, medallion engineering, Spark transformations	Relational analytics, dimensional models, SQL serving	Telemetry, logs, events, time-series analytics
Main languages	PySpark, SQL, notebooks	T-SQL	KQL
Data style	Open lake data, tables and files	Structured relational tables	High-volume event data
Common exam clue	“raw/curated files,” “Spark,” “Delta,” “engineering pipeline”	“SQL-first,” “star schema,” “warehouse,” “T-SQL”	“telemetry,” “logs,” “real-time,” “KQL,” “Eventstream”
Avoid when	Requirement is purely relational SQL warehouse serving	Requirement needs open Spark/file processing	Requirement is batch dimensional warehouse only

Dataflow Gen2 vs Notebook vs Pipeline

Requirement	Dataflow Gen2	Notebook	Pipeline
Main role	Low-code transform	Code-first transform	Orchestration/control flow
Best for	Power Query transformations, analyst-friendly ETL	PySpark/SQL transformations, complex logic, scalable processing	Scheduling, dependencies, parameters, retries, multi-step workflows
Not best for	Heavy custom code or complex distributed algorithms	Simple low-code transformations owned by business users	Complex row-by-row transformation logic by itself
Exam clue	“low-code,” “Power Query,” “business analyst can maintain”	“PySpark,” “custom logic,” “large-scale transform”	“schedule,” “trigger,” “dependency,” “retry,” “parameterize”

Shortcut vs Copy vs Mirroring

Requirement	Shortcut	Copy/ingest	Mirroring
What it does	References data in place	Physically moves data	Replicates supported operational sources
Best when	Avoid duplication; access external/internal data through OneLake	Need curated copy, transformation, isolation, or controlled landing	Need near real-time operational database replication with minimal ETL
Main trap	It does not transform or own the data	Can duplicate data and add latency	Not a generic replacement for all ETL
Exam clue	“no copy,” “single copy,” “access data where it resides”	“land data,” “transform,” “store curated version”	“replicate operational database,” “minimal ETL,” “near real-time”

Batch vs Streaming transformation

Scenario	Preferred approach	Why
Nightly load from CRM	Pipeline + Dataflow Gen2/notebook/T-SQL	Batch orchestration with scheduled dependency
Large data lake transformation	Notebook with Spark	Distributed processing and engineering flexibility
SQL dimensional load	Warehouse + T-SQL	SQL-native modeling and serving
IoT events in near real time	Eventstream + Eventhouse/KQL	Event ingestion and time-series querying
Continuous stream with custom logic	Spark structured streaming	Code-first streaming transformation
Incremental source table load	Pipeline with watermark	Avoids reprocessing all data

Security and governance selection

Requirement	Best mechanism	Avoid confusing with
Give user broad workspace collaboration	Workspace role	Item permission
Give access to one specific artifact	Item permission	Workspace Admin role
Restrict rows by user	Row-level security	Sensitivity label
Hide sensitive columns	Column-level security or masking	Workspace role
Protect/classify confidential data	Sensitivity label	RLS/CLS
Mark trusted content	Endorsement/certification	Security permission
Audit actions	Fabric audit logs	Refresh history only
Control OneLake file/table access	OneLake security	SQL-only permission

6. Architecture Patterns

Pattern 1: Enterprise medallion lakehouse

Scenario: Raw files arrive from multiple sources. Engineers need scalable transformations and curated tables for analytics.

Recommended solution:

Land raw data in a lakehouse bronze area.
Use notebooks/Spark for cleansing, deduplication, schema handling, and enrichment.
Store curated silver/gold Delta tables.
Orchestrate with pipelines.
Use deployment pipelines and Git for lifecycle.
Apply OneLake security, item permissions, labels, and audit monitoring.